bedf5497-2

ConfirmedSecurity522 words3 min decode

Coinbase AI Blocked $4.7M in Scam Transfers, Then Let One Through on a Lie

Coinbase says its AI blocked $4.7 million in transfers after 500,000 watched a deepfake giveaway, but let one $5,000 payment through when a customer lied about its purpose.

Coinbase AI blocks $4.7 million scam, but a simple lie exposes its weakness
WitnessCoinbase AI blocks $4.7 million scam, but a simple lie exposes its weaknessPexels / pixabay

Outputs

  1. Coinbase AI intercepted roughly $4.7 million in transfers tied to a deepfake giveaway scam

  2. More than 500,000 people watched the fraudulent broadcast; attempts spanned about two hours

  3. A customer's $100,000 transfer was blocked, but a second $5,000 attempt passed after a false explanation

  4. VP of product Will Megson disclosed the incident and the system's confidence-score design

  5. Coinbase is expanding AI detection models, automated reviews and scam-technique identification

Coinbase's AI fraud-prevention system intercepted roughly $4.7 million in customer transfers tied to a deepfake giveaway scam, the exchange disclosed, after more than 500,000 people watched the fraudulent broadcast and hundreds of customers attempted to send millions of dollars within about two hours.

Unlike a typical account takeover, the customers authorized the payments themselves. They believed they were joining a legitimate promotion, which removed the usual signals — unfamiliar devices, changed credentials — that fraud systems rely on.

How the interception worked

Coinbase said its AI system flagged suspicious payments in real time and engaged customers before funds left the platform, without requiring human intervention. Will Megson, Coinbase's vice president of product working on risk, outlined the architecture behind that response.

The system combines machine-learning models that identify suspicious transactions with an AI agent that examines the payment's context and questions customers about potential fraud. The agent assigns each customer response a confidence score before any decision reaches the customer.

Coinbase said the system had never encountered this specific combination of a deepfake giveaway and a live product demonstration. It still recognized familiar warning signs, including promises of guaranteed returns and irreversible payments.

What did the $100,000 case reveal?

The exchange disclosed one incident that shows the limits of its defenses. Megson said a customer first attempted to transfer $100,000 to the scammers.

The AI agent identified the transaction, explained the fraudulent giveaway and convinced the customer to abandon the payment. The customer then returned and attempted a second transfer, this time for $5,000.

When the agent questioned the transaction, the customer claimed the funds were intended for their daughter. The AI accepted the explanation and allowed the payment to proceed.

Megson acknowledged how hard it is to separate legitimate payments from fraudulent ones when customers give misleading explanations. The episode illustrates a structural constraint on automated fraud detection: an exchange can correctly identify a scam, yet customers can still complete the payment by adjusting transaction details or misrepresenting their intent.

Why authorized transfers are harder to stop

The case parallels a problem banks have faced for years with authorized push payment fraud, where victims initiate transfers themselves and insist on completing them. An AI agent that questions intent must eventually weigh a plausible answer against a suspicious transaction — and a confident lie from the account holder tips that balance.

The $5,000 loss is small next to the $4.7 million the system blocked, but it identifies the failure mode: the defense depends partly on the honesty of the person it is protecting. Confidence scores on customer responses can rank explanations, yet they cannot verify them.

What Coinbase is changing

Coinbase said it is expanding its AI-powered fraud prevention through three tracks:

  • improved detection models;
  • automated reviews of suspicious transactions;
  • faster identification of emerging scam techniques.

The exchange did not publish a timeline for those upgrades or specify how the confidence-scoring mechanism will handle deliberately false customer statements. As deepfake broadcasts grow more convincing, exchanges will need detection that holds even when the customer vouches for the transaction.

By Tom Whitfield

via cryptonews.net (Source)

More from Tom Whitfield

Tom Whitfield

Show full bio

News editor covering technology and product news in crypto and web3 industry at Mempool Brief.

1,440 articles